ANALYSIS OF SPEAKER ADAPTATION TECHNIQUES IN AUTOMATIC SPEECH RECOGNITION SYSTEMS USING DEEP NEURAL NETWORKS AND GAUSSIAN MIXTURE MODELS

The accuracy of automatic speech recognition (ASR) systems can be affected by changes in training and
testing environments. In the context of adaptation, narrowing the model-to-dataset discrepancy for a certain
speaker or channel is quite effective. Two of the most widely used ASR methods nowadays are deep neural
networks and Gaussian mixture models (GMMs). GMM-HMM has been a standard approach in ASR
systems for decades. Speaker adaption is especially helpful to AMs in this particular subgroup. Efforts have
been made to help this group in several ways. DNN-HMM AMs, on the other hand, have lately beaten
GMM-HMM models in ASR tasks. These AMs, on the other hand, frequently have to retrain their accents,
which can be difficult for them. As a result, many GMM model modification processes do not apply to
DNNs. An explanation of GMM models, as well as ways for increasing speaker adaption, is the primary
purpose of this study Domain Adaptation Challenge unsupervised domain adaptation goal data might be
collected using DNNs as well (DAC). An out-of-domain system's speaker recognition performance is
improved by more than 25 percent by using a DNN trained on data from outside the domain.